A robust adaptive beamforming method for arrays based on deep learning

Abstract

The invention discloses an array robust adaptive beamforming method based on deep learning, comprising the following steps: constructing an array element-level radar echo model based on radar system parameters and taking non-ideal factors into consideration, and simulating and generating an estimated echo covariance matrix set under a small snapshot; constructing a two-level cascade network model based on a convolutional neural network; training the two-level cascade network model by adopting a strategy of hierarchical pre-training and cascade full network fine-tuning; performing analog-to-digital conversion on the actual radar echo covariance matrix set, and then inputting a small amount of sample data of the estimated echo covariance matrix into the trained two-level cascade network model for processing to obtain an adaptive weight vector; processing the range gate data to be detected according to the obtained adaptive weight vector to achieve effective suppression of interference; performing constant false alarm processing on the result after suppressing interference and target accumulation to complete the detection of moving targets.

Description

Technical Field

[0001] The present invention relates to the technical field of radar array signal processing, and more specifically, to an array robust adaptive beamforming method based on deep learning. Background Art

[0002] In modern warfare, high-performance radar is one of the indispensable high-tech equipment, and it has already penetrated into our lives, such as civil aircraft, weather radar, reversing radar, etc. Radar uses the characteristics of electromagnetic waves propagating in a straight line in space and being scattered by objects to complete target detection and parameter estimation. Phased array radar is a widely used radar system with the characteristics of short response time and strong anti-interference ability. Unlike traditional mechanical scanning radars that rely on the rotation of the antenna itself to achieve beam scanning, phased array radars use digital systems to control the phase of each array element on the antenna array to achieve changes in beam pointing.

[0003] Digital beamforming (DBF) is an important technology unique to phased array radar. It forms a receiving beam with the maximum gain direction pointing to the direction of the desired signal through digital signal processing, thereby realizing spatial filtering and improving the performance of radar signal processing. However, the spatial electromagnetic environment in which the radar works is relatively complex. The desired signal, interference and noise usually coexist, and the desired signal is often submerged in the high-power noise suppression interference signal. In order to solve this problem, adaptive digital beamforming (ADBF) technology came into being. ADBF technology is an organic combination of adaptive signal processing technology and DBF technology. It uses an algorithm to obtain an adaptive weight vector to weight the array receiving signal, thereby controlling the beam shape of the phased array radar. The weighted beam forms the maximum gain in the direction of the desired signal and forms a deep null in the direction of the interference position, which can greatly improve the output signal-to-interference-to-noise ratio (SINR). The clutter covariance matrix required for the calculation of the optimal weight vector is unknown, so it needs to be obtained by maximum likelihood estimation of the clutter data (training samples). This means that enough training samples are needed to ensure the effectiveness of ADBF.

[0004] When the prior information of the phased array radar is accurate and error-free, the performance of ADBF is extremely superior. However, in actual scenarios, there are often various non-ideal factors, such as inaccurate angle estimation of the desired signal, phase, amplitude and position errors between the elements of the phased array, etc., which lead to the performance of traditional ADBF technology being impaired. At the same time, the power or waveform of the new active electronic interference signal changes rapidly in the time domain, that is, the signal has serious non-stationary characteristics in the time domain, resulting in a serious shortage of stationary training samples for estimating the echo covariance matrix, which causes the performance of the traditional ADBF method to drop sharply. Therefore, it is of great significance to study the formation of robust adaptive beamforming (RBF) under non-ideal conditions.

[0005] At present, RBF can be mainly divided into three categories. The first category is diagonal loading or regularization technology, the core idea of ​​which is to add a unit matrix to the estimated covariance matrix multiplied by the diagonal loading factor. In a Gaussian environment, the covariance matrix of the noise is a unit matrix, so diagonal loading is equivalent to uniformly correcting the noise power to the same level, thereby eliminating the influence of noise divergence on the signal and improving the robustness of ADBF. However, the diagonal loading technology has a very significant disadvantage. Its diagonal loading coefficient needs to be set manually using prior information, and there is no deterministic criterion to guide it. When the diagonal loading factor is too small, the improvement on robustness is weak, and when the diagonal loading factor is too large, the interference suppression effect will be reduced. The second category is the RBF method based on feature subspace projection. The core of this method is to divide the entire data space into two orthogonal subspaces, namely the signal plus interference subspace and the noise subspace, and then project the desired signal steering vector into the noise subspace and the signal plus interference subspace. Since the two subspaces are orthogonal, the error component cannot affect the final adaptive weight vector. However, in the case of low signal-to-noise ratio, the difference between the eigenvalues ​​of noise and signal or interference is not significant, and it is difficult to accurately divide the above two orthogonal subspaces, which ultimately limits the performance of this type of method. The third type is RBF based on convex optimization theory, among which the most representative algorithm is the worst performance optimization algorithm. This algorithm models the true desired signal steering vector as the sum of an estimated desired signal steering vector and an unknown error vector, assuming that the error vector is in a spherical uncertainty set, and constrains all steering vectors to pass through the beamformer without distortion, thereby solving the optimal weight vector. This type of method can be regarded as an extension of the diagonal loading technology, and the diagonal loading amount is given by solving the optimization problem. However, this type of method has a strong dependence on the model, and the accuracy of its solution is directly related to the selection of the model. When the signal parameters change drastically, serious performance loss may occur.

[0006] Therefore, to effectively suppress interference under actual non-ideal conditions, a data-driven ADBF method with small sample requirements and robustness to various errors is needed. Summary of the invention

[0007] In order to solve the above problems of the prior art, the present invention provides an array robust adaptive beamforming method based on deep learning.

[0008] In order to achieve the above-mentioned purpose of the present invention, the technical scheme adopted is as follows:

[0009] A method for array robust adaptive beamforming based on deep learning, the method comprising the following steps:

[0010] S1: Based on radar system parameters and considering non-ideal factors, an element-level radar echo model is constructed to simulate and generate an estimated echo covariance matrix set under small snapshots;

[0011] S2: Construct a two-level cascade network model based on a convolutional neural network; use the estimated echo covariance matrix set under the small snapshot as the input of the first-level network, and use the known echo covariance matrix output by the first-level network as the label data set of the first-level network; at the same time, use the known echo covariance matrix as the input of the second-level network, and use the minimum variance distortion-free weight coefficient corresponding to the known echo covariance matrix as the label data set of the second-level network;

[0012] S3: The two-level cascade network model is trained using the strategy of hierarchical pre-training and cascade full network fine-tuning;

[0013] S4: After performing analog-to-digital conversion on the actual radar echo covariance matrix set, a small amount of sample data is used to estimate the echo covariance matrix and input it into the trained two-level cascade network model for processing to obtain an adaptive weight vector;

[0014] S5: Processing the range gate data to be detected according to the obtained adaptive weight vector to achieve effective suppression of interference;

[0015] S6: Perform constant false alarm processing on the results after interference suppression and target accumulation to complete the detection of moving targets.

[0016] Preferably, the non-ideal factors include array element amplitude / phase error, interference number, interference direction and interference-to-noise ratio.

[0017] Furthermore, when the estimated echo covariance matrix set under small snapshots is generated by simulation,

[0018] Considering that the amplitude and phase errors of the array elements satisfy the complex Gaussian distribution, the variance selection is limited to the actual error range of the current array antenna, and the variance is randomly set in the closed interval [0, 0.05];

[0019] The number of interferences is randomly set and is within 3;

[0020] The interference-to-noise ratio setting range is set according to the actual combat scenario, and the value is in the range of 25dB to 35dB;

[0021] The interference direction is randomly selected from the range of [-70°, 10°] ∪ [10°, 70°] with an interval of 1°.

[0022] Preferably, the minimum variance distortion-free weight coefficient is obtained based on the following formula:

[0023] w=μR -1 s (1)

[0024] in, is the normalization constant, (·) H represents the conjugate transpose, (·)-1 represents the inverse of the matrix, and s represents the array target spatial domain steering vector.

[0025] Preferably, step S3, the specific training is as follows: first, the first-level network and the second-level network are trained separately, and then the cascaded second-level cascade network model is fine-tuned; during the training process, the Adam optimizer and gradual warm up are used to dynamically adjust the learning rate to improve the convergence rate.

[0026] Preferably, the first-level network is used to complete the estimation of the estimated echo covariance matrix set, and the second-level network implements the inversion of the estimated echo covariance matrix and the product operation process with the target-oriented vector to obtain an adaptive weight vector.

[0027] Furthermore, the specific structure of the first-level network includes: an input layer with an input size of 16*16, a convolution block with 4 channels, two convolution blocks with 16 channels, two residual blocks with 16 channels, and a 1*1 convolution layer connected in sequence.

[0028] Furthermore, the specific structure of the second-level network includes: an intermediate input layer with an input size of 16*16, a convolution block with 4 channels, a convolution block with 16 channels, a 2*2 downsampling layer, two residual blocks with 16 channels, a linear flattening layer and a linear output layer with an output scale of 32*1 connected in sequence.

[0029] A computer device comprises a memory, a processor and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the steps of the array robust adaptive beamforming method based on deep learning are implemented.

[0030] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of the array robust adaptive beamforming method based on deep learning.

[0031] The beneficial effects of the present invention are as follows:

[0032] The present invention takes advantage of the current deep learning technology that can achieve complex nonlinear mapping. The estimated echo covariance matrix set, known echo covariance matrix and accurate adaptive weight vector obtained by phased array multi-channel array radar echo modeling simulation are used to train the constructed two-level cascade network model. The strategy of segmented pre-training and cascade full network fine-tuning is adopted during training, and the network learning rate is dynamically adjusted to improve the convergence performance, so that it can achieve end-to-end RBF function with very few sample requirements. After the network training is completed, the echo covariance matrix estimated by a small number of samples in the actual radar observation data is input into the neural network for processing, and then the accurate adaptive weight vector is obtained at the output end, thereby achieving effective suppression of interference in the range gate data to be detected. The scheme of the present invention is suitable for adaptive beamforming of arrays, and has the characteristics of low sample requirements, robust errors and simple engineering implementation.

[0033] It should be noted that in the invented method, the array element level radar echo modeling, the construction of the two-level cascade network model and the two-level cascade network model training are all offline processing, while the subsequent measured data neural network processing, the construction of the adaptive weight vector, the processing of the range gate data to be detected and the constant false alarm detection part are online processing, so the required amount of calculation is relatively small, which is more suitable for practical engineering applications. At the same time, compared with the previous RBF method, the invented method can show better interference suppression performance when the number of snapshots is very small, and it has a certain robustness to amplitude and phase errors. This method is suitable for multi-channel digital array systems, such as phased array radars and smart antennas. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 It is a flowchart of the steps of the array robust adaptive beamforming method based on deep learning described in the present invention.

[0035] Figure 2 It is a schematic diagram of the structure of the first-level network.

[0036] Figure 3 It is a schematic diagram of the structure of the second-level network.

[0037] Figure 4 It is a schematic diagram of the structure of the convolution block and the residual block. DETAILED DESCRIPTION

[0038] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments.

[0039] Example 1

[0040] like Figure 1 A method for array robust adaptive beamforming based on deep learning, the method comprising the following steps:

[0041] S1: Based on radar system parameters and considering non-ideal factors, an element-level radar echo model is constructed to simulate and generate an estimated echo covariance matrix set under small snapshots;

[0042] S2: Construct a two-level cascade network model based on a convolutional neural network; use the estimated echo covariance matrix set under the small snapshot as the input of the first-level network, and use the known echo covariance matrix output by the first-level network as the label data set of the first-level network; at the same time, use the known echo covariance matrix as the input of the second-level network, and use the minimum variance distortion-free weight coefficient corresponding to the known echo covariance matrix as the label data set of the second-level network;

[0043] S3: The two-level cascade network model is trained using the strategy of hierarchical pre-training and cascade full network fine-tuning;

[0044] S4: After performing analog-to-digital conversion on the actual radar echo covariance matrix set, a small amount of sample data is used to estimate the echo covariance matrix and input it into the trained two-level cascade network model for processing to obtain an adaptive weight vector;

[0045] S5: Processing the range gate data to be detected according to the obtained adaptive weight vector to achieve effective suppression of interference;

[0046] S6: Perform constant false alarm processing on the results after interference suppression and target accumulation to complete the detection of moving targets.

[0047] The present invention takes advantage of the current deep learning technology that can achieve complex nonlinear mapping. The estimated echo covariance matrix set, known echo covariance matrix and accurate adaptive weight vector obtained by phased array multi-channel array radar echo modeling simulation are used to train the constructed two-level cascade network model. The strategy of segmented pre-training and cascade full network fine-tuning is adopted during training, and the network learning rate is dynamically adjusted to improve the convergence performance, so that it can achieve end-to-end RBF function with very few sample requirements. After the network training is completed, the echo covariance matrix estimated by a small number of samples in the actual radar observation data is input into the neural network for processing, and then the accurate adaptive weight vector is obtained at the output end, thereby achieving effective suppression of interference in the range gate data to be detected. The scheme of the present invention is suitable for adaptive beamforming of arrays, and has the characteristics of low sample requirements, robust errors and simple engineering implementation.

[0048] It should be noted that in the invented method, the array element level radar echo modeling, the construction of the two-level cascade network model and the two-level cascade network model training are all offline processing, while the subsequent measured data neural network processing, the construction of the adaptive weight vector, the processing of the range gate data to be detected and the constant false alarm detection part are online processing, so the required amount of calculation is relatively small, which is more suitable for practical engineering applications. At the same time, compared with the previous RBF method, the invented method can show better interference suppression performance when the number of snapshots is very small, and it has a certain robustness to amplitude and phase errors. This method is suitable for multi-channel digital array systems, such as phased array radars and smart antennas.

[0049] In a specific embodiment, the non-ideal factors include array element amplitude / phase error, interference number, interference direction and interference-to-noise ratio.

[0050] In this embodiment, when simulating and generating a set of estimated echo covariance matrices considering different array element errors, different interference-to-noise ratios, different interference directions, and different interference numbers under a small snapshot,

[0051] Considering that the amplitude and phase errors of the array elements satisfy the complex Gaussian distribution, the variance selection is limited to the actual error range of the current array antenna, and the variance is randomly set in the closed interval [0, 0.05];

[0052] The number of interferences is randomly set to be within 3, that is, the number of interferences is randomly selected from 0, 1, 2, and 3;

[0053] The interference-to-noise ratio setting range is set according to the actual combat scenario, and the value is in the range of 25dB to 35dB;

[0054] The interference direction is randomly selected from the range of [-70°, 10°] ∪ [10°, 70°] with an interval of 1°.

[0055] In a specific embodiment, the organization form of the input data set and the label data set is related to the cascade characteristics of the network. This embodiment uses the estimated echo covariance matrix set under adjacent small snapshots of the range gate to be detected as the input data set of the first-level network, and the known covariance matrix corresponding to the range gate is used as the label data set of the first-level network; at the same time, the minimum variance distortion-free weight coefficient (MVDR) corresponding to the known covariance matrix is ​​used as the label data set of the second-level network.

[0056] The minimum variance distortion-free weight coefficient is obtained based on the following formula:

[0057] w=μR -1 s (1)

[0058] in, is the normalization constant, (·) H represents the conjugate transpose, (·)-1 represents the inverse of the matrix, and s represents the array target spatial domain steering vector.

[0059] In a specific embodiment, step S3 is specifically trained as follows: first, the first-level network and the second-level network are trained separately, and then the cascaded second-level cascade network model is fine-tuned; during the training process, the Adam optimizer and gradual warm up are used to dynamically adjust the learning rate to improve the convergence rate.

[0060] In a specific embodiment, the first-level network is used to complete the estimation of the estimated echo covariance matrix set, and the second-level network implements the inversion of the estimated echo covariance matrix and the product operation process with the target steering vector to obtain the adaptive weight vector.

[0061] The first-level network and the second-level network are both convolutional neural networks, and the two networks are in a cascade relationship, which can be trained separately, and finally the whole network is fine-tuned to finally realize the nonlinear mapping from the echo covariance matrix estimated by the input small snapshot to the minimum variance distortion-free weight coefficient.

[0062] In a specific embodiment, Figure 2 As shown, the specific structure of the first-level network includes: an input layer with an input size of 16*16, a convolution block with 4 channels, two convolution blocks with 16 channels, two residual blocks with 16 channels, and a 1*1 convolution layer connected in sequence.

[0063] In a specific embodiment, Figure 3 As shown, the specific structure of the second-level network includes: an intermediate input layer with an input size of 16*16, a convolution block with 4 channels, a convolution block with 16 channels, a 2*2 downsampling layer, two residual blocks with 16 channels, a linear flattening layer and a linear output layer with an output scale of 32*1 are sequentially connected.

[0064] like Figure 4 As shown in the figure, the convolution block is composed of a 3*3 convolution layer, a batch normalization layer and an activation function layer connected in sequence; the residual block adds a residual connection on the basis of the convolution block; the activation function selects the ReLU function:

[0065] ReLU(x)=max{x,0} (2)

[0066] Where max{·} is the maximum value function, which takes the maximum value between x and 0.

[0067] In a specific embodiment, a two-level cascade network model is trained based on input and output data sets, the training batch size is set to 256, the two-level pre-training and the final unified fine-tuning are trained 200 times, and the Adam optimizer and gradual warm up are used to dynamically adjust the learning rate during the training process. The initial learning rate is 0.001, the warm up step size is 5, and the interval is 0.002.

[0068] In a specific embodiment, a small number of adjacent samples are taken from the radar actual echo covariance matrix set to estimate the echo covariance matrix, and the typical value of the number of samples is generally 4 to 8. In this embodiment, the actual echo covariance matrix sets of N radars are respectively converted into analog-to-digital to realize the digitization of the received data and store it.

[0069] In this embodiment, four range gate data adjacent to the range gate to be detected are taken as training samples, and a covariance matrix is ​​obtained through likelihood estimation. The estimated echo covariance matrix is ​​input into a two-level cascade network model, and an adaptive weight vector is obtained at the output end of the two-level cascade network model.

[0070] The obtained adaptive weight vector is used to act on the echo range gate data to achieve interference suppression in the data and coherent accumulation of the target. Specifically, the output adaptive weight vector is conjugate multiplied with the array received range gate data to be detected to achieve interference suppression and coherent accumulation of the target.

[0071] Finally, the adaptively processed data is subjected to constant false alarm processing to complete the detection of radar moving targets. The constant false alarm processing can be processed by selecting an existing classic constant false alarm processing algorithm.

[0072] Example 2

[0073] A computer device comprises a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the steps of the array robust adaptive beamforming method based on deep learning as described in Example 1 are implemented.

[0074] Among them, the memory and the processor are connected in a bus manner, and the bus may include any number of interconnected buses and bridges, and the bus connects various circuits of one or more processors and memories together. The bus can also connect various other circuits such as peripherals, voltage regulators, and power management circuits, which are well known in the art and are therefore not further described herein. The bus interface provides an interface between the bus and the transceiver. The transceiver can be one element or multiple elements, such as multiple receivers and transmitters, providing a unit for communicating with various other devices on a transmission medium. The data processed by the processor is transmitted on a wireless medium via an antenna, and further, the antenna also receives data and transmits the data to the processor.

[0075] Example 3

[0076] A computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, the steps of the array robust adaptive beamforming method based on deep learning as described in Example 1 are implemented.

[0077] That is, those skilled in the art can understand that all or part of the steps in the above-mentioned embodiment method can be completed by instructing the relevant hardware through a program, and the program is stored in a storage medium, including several instructions to enable a device (which can be a single-chip microcomputer, chip, etc.) or a processor to execute all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk and other media that can store program codes.

[0078] Obviously, the above embodiments of the present invention are only examples for clearly explaining the present invention, and are not intended to limit the implementation methods of the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention shall be included in the protection scope of the claims of the present invention.

Claims

1. A method for robust adaptive beamforming of an array based on deep learning, characterized in that: The method comprises the following steps: S1: Based on radar system parameters and considering non-ideal factors, an element-level radar echo model is constructed to simulate and generate an estimated echo covariance matrix set under small snapshots; S2: A two-level cascade network model is constructed based on a convolutional neural network; the estimated echo covariance matrix set under the small snapshot is used as the input of the first-level network, and the known echo covariance matrix output by the first-level network is used as the output label data set of the first-level network; at the same time, the known echo covariance matrix is ​​used as the input of the second-level network, and the minimum variance distortion-free weight coefficient corresponding to the known echo covariance matrix is ​​used as the output label data set of the second-level network; the first-level network is used to complete the estimation of the estimated echo covariance matrix set, and the second-level network realizes the inversion of the estimated echo covariance matrix and the product operation process with the target-oriented vector to obtain an adaptive weight vector; S3: The two-level cascade network model is trained using the strategy of hierarchical pre-training and cascade full network fine-tuning; S4: After performing analog-to-digital conversion on the actual radar echo covariance matrix set, a small amount of sample data is used to estimate the echo covariance matrix and input it into the trained two-level cascade network model for processing to obtain an adaptive weight vector; S5: Processing the range gate data to be detected according to the obtained adaptive weight vector to achieve effective suppression of interference; S6: Perform constant false alarm processing on the results after interference suppression and target accumulation to complete the detection of moving targets.

2. The array robust adaptive beamforming method based on deep learning according to claim 1, characterized in that: The non-ideal factors include array element amplitude / phase error, interference number, interference direction and interference-to-noise ratio.

3. The array robust adaptive beamforming method based on deep learning according to claim 2, characterized in that: When the simulation generates the estimated echo covariance matrix set under small snapshots, Considering that the amplitude and phase errors of the array elements satisfy the complex Gaussian distribution, the variance selection is limited to the actual error range of the current array antenna, and the variance is randomly set in the closed interval [0, 0.05]; The number of interferences is randomly set and is within 3; The interference-to-noise ratio setting range is set according to the actual combat scenario, and the value is in the range of 25dB to 35dB; The interference direction is randomly selected from the range of [-70°, 10°] ∪ [10°, 70°] with an interval of 1°.

4. The array robust adaptive beamforming method based on deep learning according to claim 1, characterized in that: The minimum variance distortion-free weight coefficient is obtained based on the following formula: w=μR -1 s (1) in, is the normalization constant, (·) H represents the conjugate transpose, (·) -1 represents the inverse of the matrix, and s represents the array target spatial domain steering vector.

5. The array robust adaptive beamforming method based on deep learning according to claim 1, characterized in that: Step S3, the specific training is as follows: first, the first-level network and the second-level network are trained separately, and then the cascaded second-level cascade network model is fine-tuned; during the training process, the Adam optimizer and gradual warm up are used to dynamically adjust the learning rate to improve the convergence rate.

6. The array robust adaptive beamforming method based on deep learning according to claim 1, characterized in that: The specific structure of the first-level network includes: an input layer with an input size of 16*16, a convolution block with 4 channels, two convolution blocks with 16 channels, two residual blocks with 16 channels, and a 1*1 convolution layer connected in sequence.

7. The array robust adaptive beamforming method based on deep learning according to claim 1, characterized in that: The specific structure of the second-level network includes: an intermediate input layer with an input size of 16*16, a convolution block with 4 channels, a convolution block with 16 channels, a 2*2 downsampling layer, two residual blocks with 16 channels, a linear flattening layer and a linear output layer with an output scale of 32*1 connected in sequence.

8. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 7 are implemented.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.

Images